Repository navigation
Automation v2: ecosystem health, evidence, and guarded OpenAI review - #6
Conversation
There was a problem hiding this comment.
⚠️ Not ready to approve
The OpenAI evidence workflow currently allows evidence-path access to .git/config (credential exfiltration risk) and the analyzer’s Responses API output parsing is brittle and likely to fail against valid responses.
Pull request overview
This PR introduces “Automation v2” for the MelodicBloom control plane: a portfolio-wide contract for deterministic Node verification (with durable evidence receipts) and an optional, guarded OpenAI evidence review workflow that operates only on bounded, caller-supplied evidence.
Changes:
- Add reusable workflows for deterministic Node verification v2 and guarded OpenAI evidence review (with artifact receipts and concurrency/timeout).
- Add the OpenAI analyzer script and a corresponding JSON schema file, plus caller examples.
- Add documentation for the v2 automation model, OpenAI evidence envelope, monetization governance, and an Actions export audit record.
File summaries
| File | Description |
|---|---|
scripts/automation/openai-evidence-review.mjs |
New analyzer that sends bounded evidence to OpenAI Responses API and writes analysis.json + provenance receipts. |
schemas/intelligence/ecosystem-recommendation.schema.json |
JSON Schema describing the structured OpenAI output envelope. |
examples/callers/opportunity-studio.yml |
Example caller for governed opportunity modes (placeholder SHA to be replaced). |
examples/callers/openai-health-review.yml |
Example caller for health review mode (placeholder SHA to be replaced). |
examples/callers/node-web-verify-v2.yml |
Example caller for deterministic Node verification v2 (placeholder SHA to be replaced). |
docs/monetization/OPPORTUNITY_GOVERNANCE.md |
Governance rules for conversion/affiliate/ads/generative work and approval boundaries. |
docs/intelligence/OPENAI_EVIDENCE_REVIEW_V2.md |
Documentation for evidence envelope, prompt-injection boundary, and prohibited actions. |
docs/automation/ECOSYSTEM_AUTOMATION_V2.md |
Overall v2 contract and rollout sequence for ecosystem automation. |
docs/automation/ACTIONS_EXPORT_AUDIT_2026-07-28.md |
Recorded audit results and prioritized repair queue from Actions exports. |
.github/workflows/reusable-openai-evidence-review.yml |
Reusable workflow to validate evidence, optionally call the analyzer, and retain artifacts. |
.github/workflows/reusable-node-verify-v2.yml |
Reusable workflow implementing frozen-install deterministic checks and evidence receipts. |
.github/workflows/control-plane-qa.yml |
Control-plane QA workflow to validate syntax and enforce safety invariants. |
Review details
Comments suppressed due to low confidence (1)
.github/workflows/reusable-openai-evidence-review.yml:99
evidence-pathcan currently reference any file under the checkout (includingsource/.git/config), and the credential scan only checks a few token patterns. This creates a realistic path to exfiltrate credentials (e.g., checkout tokens) into the evidence sent to OpenAI. Add a path allow/deny check (no.., no absolute paths, forbid.git/) and broaden the credential regex to catch common GitHub token formats andx-access-token:.
set -euo pipefail
mkdir -p "$OUTPUT_DIR"
test -f "$EVIDENCE_PATH"
bytes=$(wc -c < "$EVIDENCE_PATH")
if (( bytes == 0 || bytes > MAX_INPUT_BYTES )); then
- Files reviewed: 12/12 changed files
- Comments generated: 4
- Review effort level: Low
Note
Your feedback helps us improve the quality of this feature.
Please use 👍 or 👎 to tell us whether this assessment is correct.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| - name: Checkout caller source revision | ||
| uses: actions/checkout@v4 | ||
| with: | ||
| path: source |
| const redact = (value) => String(value ?? "Unknown failure") | ||
| .replace(/sk-[A-Za-z0-9_-]+/g, "[REDACTED_OPENAI_KEY]") | ||
| .replace(/github_pat_[A-Za-z0-9_]+/g, "[REDACTED_GITHUB_TOKEN]") | ||
| .replace(/Bearer\s+[^\s]+/gi, "Bearer [REDACTED]") | ||
| .slice(0, 2000); |
| "title": "Ecosystem Evidence Review", | ||
| "type": "object", | ||
| "additionalProperties": false, | ||
| "required": [ |
| const outputText = payload.output | ||
| ?.flatMap((item) => item.type === "message" ? item.content ?? [] : []) | ||
| .find((content) => content.type === "output_text") | ||
| ?.text; | ||
| if (!outputText) throw new Error("OpenAI response did not contain output_text"); | ||
| const analysis = JSON.parse(outputText); |
Purpose
Evolve the merged MelodicBloom automation control plane into a portfolio-wide v2 contract for reproducible verification, durable failure evidence, and optional OpenAI-assisted analysis of bounded evidence.
Evidence from the uploaded Actions exports
The three exports cover 37 workflows, 54 jobs, and 17 repositories. The highest-priority signals are:
build.ymlfailed 59/59 and currently combines non-frozen installation, lockfile mutation, commit, and push inside normal CI.npm installoperations and no committedpackage-lock.jsonwas found.What changed
reusable-node-verify-v2.ymlreusable-openai-evidence-review.ymlOPENAI_API_KEY; missing key produces an explicit skipped receiptpull_request_targetstore: falseSecurity and authority boundary
OPENAI_API_KEYin GitHub Actions secretsQA performed before upload
node --checkNot included
Next gated phase
After review and merge, pin canary callers to the immutable merge SHA and repair repositories in this order: Flore Orthography, Wildfire Iridescent, Nacre Orchid Atelier, Omni-Loom Case Study, AETHER, Mochi UI, and ChromaFlora Prime. Require seven consecutive green deterministic runs before promoting a new check to required status.
Do not merge without Jennipher approval.